Trust as Triangulation: From E-E-A-T to a Three-Evaluator, Nine-Property Credibility Model for Human and AI-Mediated Evaluation 

Trust as Triangulation: From E-E-A-T to a Three-Evaluator, Nine-Property Credibility Model for Human and AI-Mediated Evaluation 

Jason Barnard Kalicube®, Aubais, France ORCID: 0009-0000-0353-6642

Status: Working paper, v1.4 (June 2026) Identifier: TKF-14-20647932, where 20647932 is the Zenodo concept DOI identifier Deposit: Zenodo, concept DOI 10.5281/zenodo.20647932 (cite all versions); version DOI 10.5281/zenodo.20647933 (v1.4) License: CC BY 4.0 International Citation: Barnard, J. (2026). Trust as Triangulation: From E-E-A-T to a Three-Evaluator, Nine-Property Credibility Model for Human and AI-Mediated Evaluation. Zenodo. https://doi.org/10.5281/zenodo.20647932 (cite as TKF-14-20647932)

Abstract

Trust evaluation in contemporary information environments is increasingly distributed across human readers, domain peers, institutions, platforms, search engines, and AI systems. Existing operational frameworks for trust assessment, including Google’s E-E-A-T and the author’s NEEATT extension, identify important trust-relevant properties and present them largely as attribute lists that are useful descriptively but less precise diagnostically. This paper argues that the operational limitation of such frameworks is not primarily the selection of properties, but the absence of evaluator specificity: they do not consistently distinguish between properties demonstrated by the assessed entity, properties conferred by domain peers, and properties constituted through audience reception or audience addressability.

The paper proposes a three-evaluator, nine-property model of trust evaluation. The model distinguishes three constitutive evaluator positions (constitutive meaning the evaluator whose judgement makes the property real in this model): Self, Peers, and Audience. Each position is associated with three operationally distinct properties: Transparency, Experience, and Expertise on the Self axis; Authority, Notability, and Corroboration on the Peers axis; and Trustworthiness, Celebrity (meaning cross-audience recognition), and Reach on the Audience axis. The central conceptual contribution is the decomposition of the Peers axis into three distinct behaviours: deference, reference, and agreement. This distinction separates Authority from Notability and Corroboration, enabling diagnoses that are unavailable when these properties are collapsed into a single notion of “authoritativeness.”

The paper further argues that the model is applicable to machine-mediated information environments because AI systems are constrained by human-produced training corpora, retrieval corpora, and human-feedback procedures. This claim, articulated as the Mirror Principle, is presented as a falsifiable expectation rather than a claim about machine cognition. The paper concludes by identifying Corroboration as a strategically compounding lever for trust development once minimum thresholds in the relevant prerequisite cells are in place, while distinguishing Corroboration from consensus, citation, and repetition.

Throughout, credibility names the layer all of these frameworks address: whether an entity should be believed, relied on, and recommended. E-E-A-T, NEEATT, and the proposed grid are alternative decompositions of that single layer. The grid is the author’s proposed decomposition, designed as a complete operational specification under the three-actor assumption; the category claim is the durable one, and it survives any future decomposition.

Keywords: credibility, trust evaluation, source credibility, E-E-A-T, NEEATT, social epistemology, computational trust, authority, notability, corroboration, AI search, retrieval-augmented generation

1. Introduction

Trust evaluation is no longer performed only by individual human readers deciding whether a source is credible. A claim made by a person, brand, institution, or publication is now evaluated across a distributed information environment. Human audiences assess whether to believe it or act on it. Domain peers cite it, contest it, ignore it, or build upon it. Platforms and search engines decide whether to surface it. Generative AI systems may retrieve it, repeat it, attribute it, synthesise it with other sources, or exclude it from an answer. Trust therefore travels through multiple evaluator positions before it reaches the person or system that acts on it.

Most operational frameworks used in search, content strategy, and digital reputation do not fully account for this distribution of evaluation. Google’s E-E-A-T framework (Google, 2023) identifies Experience, Expertise, Authoritativeness, and Trustworthiness as important dimensions of quality assessment. NEEATT, the present author’s extension of E-E-A-T with Transparency contributed by Jarno van Driel (Barnard, TKF-7), adds Notability and Transparency to that set. These frameworks name properties that matter, but they do not consistently specify which evaluator position is constitutive of each property. As a result, they are useful as descriptive checklists and less precise as diagnostic tools. An entity may be experienced but not expert, expert but not authoritative, authoritative but not notable, notable but not corroborated, or trusted by a narrow audience while remaining unreachable by a wider one.

1.1 The argument

This paper argues that trust evaluation becomes operationally precise when properties are assigned to the evaluator position capable of constituting them. The proposed model distinguishes three such positions: Self, Peers, and Audience. The Self can disclose, demonstrate, and document properties such as Transparency, Experience, and Expertise. Peers can confer Authority, constitute Notability, and produce Corroboration through deference, reference, and agreement. Audiences can constitute audience-perceived Trustworthiness, Celebrity, and Reach through confidence, recognition, and addressability. In this model, Trustworthiness names audience-side confidence in the entity, not the full moral or organisational property of being trustworthy, and the Audience axis includes reception conditions as well as judgements. The actor assigned to a property is not necessarily the only actor that can observe evidence of that property. Rather, it is the actor whose judgement is constitutive of the property within the model.

The paper’s central conceptual contribution is the decomposition of the Peers axis. Existing trust frameworks often treat authority as a broad aggregate of reputation, recognition, citation, visibility, and agreement. This paper separates those behaviours into three distinct properties. Authority is deference: peers treat the entity’s judgement as carrying weight on field-relevant questions. Notability is reference: peers name or discuss the entity as salient within the domain. Corroboration is agreement: independent voices align with the entity’s claims, framings, or findings. These properties can move independently. A figure may be widely discussed without being deferred to, deferred to privately and rarely named publicly, or supported by diffuse agreement without being personally notable. The model’s diagnostic value lies in making those differences visible.

The second contribution is operational. A nine-cell trust profile makes it possible to identify the bottleneck cell in a given entity’s trust position and to name the actor whose behaviour must change. This is the difference between a checklist and a diagnostic framework. A practitioner with strong Experience and Expertise but weak Corroboration does not need the same intervention as a public figure with high Reach and Celebrity but low Authority. A brand with high Notability and low Trustworthiness requires a different intervention from a specialist with high Authority and low Reach. The model is designed to distinguish these cases and to connect each diagnosis to an intervention path.

The third contribution concerns machine-mediated evaluation. AI systems do not need to “think like humans” for human trust structures to shape their outputs. Their behaviour is constrained by human-authored training corpora, human-produced retrieval corpora, and human-feedback procedures (Bender et al., 2021; Ouyang et al., 2022; Lewis et al., 2020). The Mirror Principle, as formulated in this paper, states that machine evaluators tend to reproduce the broad trust-signal structure embedded in the human information environments from which they learn and retrieve. This is not an identity claim about machine cognition. It is a falsifiable expectation about the relationship between human-produced trust signals and machine-mediated surfacing, citation, recommendation, and exclusion.

1.2 Method and scope

This is a conceptual model paper. Its scope is the structure of trust evaluation across human and machine evaluators, and the operational diagnostic that the proposed structure makes available. The paper does not empirically validate the model. Empirical validation would require, at minimum: longitudinal case studies that score entities across the nine cells over time and observe trust outcomes; controlled studies that compare diagnoses produced by E-E-A-T, NEEATT, and the nine-cell grid against independent ground truth; and machine-evaluator experiments that test the Mirror Principle’s predictions across training corpora, retrieval configurations, and feedback regimes. Section 10 names limitations and the empirical work that would address them. The paper’s claim, accordingly, is conceptual: the proposed grid offers a bounded operational specification under the paper’s three-actor assumption, and the specification supports diagnostic uses that flat-list frameworks do not. A terminological note bounds the vocabulary: terms such as Authority, Notability, Corroboration, Trustworthiness, Celebrity, and Reach are used operationally, not as exhaustive definitions drawn from sociology, psychology, philosophy, or information retrieval. The model’s claim is diagnostic, and assigning each property to its constitutive evaluator position improves intervention design. This paper does not claim to exhaust all philosophical, sociological, or organisational theories of trust. It offers an operational diagnostic model for public information environments: trust-relevant signals become actionable when assigned to the evaluator position that constitutes them.

1.3 Research lineage

This paper sits within an applied research programme on entity trust and AI-mediated recommendation that includes AI-Era Business Engineering, the programme’s integrating framework statement (Barnard, TKF-10), and the earlier NEEATT extension of E-E-A-T (Barnard, TKF-7), in which Notability was added by Jason Barnard and Transparency by Jarno van Driel. Notability was added because the predecessor list contained no property for recognition: a framework with no cell for recognition cannot diagnose the difference between being known and being deferred to. NEEATT extended E-E-A-T, and the nine-cell grid extends the extension: four named properties became six, and six become nine once each property is assigned to the evaluator position constitutive of it. The lineage is relevant because it explains the operational problem from which the model emerged. Within the programme’s Understandability, Credibility, Deliverability model, the nine-cell grid operationalises the Credibility layer: E-E-A-T, NEEATT, and the grid are, in those terms, credibility instruments. The category claim is framework-agnostic: practitioners may prefer other property lists, and any such list is a decomposition of the same Credibility layer. The durable claim is the layer; the grid is the author’s proposed decomposition of it. The present paper, however, treats the model as a conceptual framework in its own right. Its claim is not that E-E-A-T is wrong. The claim is that E-E-A-T identifies part of a larger evaluator-specific structure, and the proposed nine-property grid preserves the useful intuitions of predecessor frameworks while specifying the evaluator positions those frameworks leave implicit.

1.4 Structure of the paper

Section 2 reviews the relevant literature in source credibility, social epistemology, authority, computational trust, AI-mediated evaluation, and search quality. Section 3 introduces the conflation problem through the distinction between Experience and Expertise. Section 4 defines the three evaluator positions. Section 5 presents the nine-property grid. Section 6 develops the peer-axis decomposition as the paper’s central conceptual contribution. Section 7 formulates the Mirror Principle for machine-mediated evaluation. Section 8 distinguishes Corroboration from consensus, citation, and repetition. Section 9 applies the model to stylised diagnostic cases. Section 10 identifies limitations and directions for empirical validation.

2. Literature review

The model draws on six literatures, none of which provides an integrated operational grid that separates evaluator positions while naming an operational set of trust properties.

2.1 Source credibility and persuasion

The classical source credibility tradition, established by Hovland, Janis, and Kelley (1953), identifies expertise and trustworthiness as the two principal dimensions on which receivers evaluate a source’s credibility. Fogg (2003) extended the model to web-based credibility evaluation through prominence-interpretation theory, distinguishing between the prominence of a credibility signal and the receiver’s interpretation of it. Metzger (2007) catalogued the credibility assessment heuristics users actually deploy when judging online information. Sundar (2008) brought interface-level factors (Modality, Agency, Interactivity, Navigability) into credibility theory through the MAIN model. Hilligoss and Rieh (2008) proposed a unified credibility assessment framework that organises the field by linking judgements at construct level, heuristic level, and interaction level. Rieh (2002) examined cognitive authority and information quality judgements specifically in web-based information seeking, providing direct empirical grounding for the Authority property as developed here. This tradition has been productive for understanding how individual users judge sources. It has not produced a framework that separates user-level evaluation from peer-level evaluation from institutional-level evaluation in operational terms. In organisational research, Mayer, Davis, and Schoorman (1995) model trustworthiness through the trustee-side qualities of ability, benevolence, and integrity, a model whose durability Schoorman, Mayer, and Davis (2007) reaffirm. Whereas organisational trust models analyse the qualities that make a trustee worthy of trust, the present model asks which evaluator positions constitute the observable trust profile of an entity in public information environments.

2.2 Social epistemology and testimony

The philosophical literature on testimony distinguishes carefully between what an individual knows, what a community recognises as known, and what testimony is competent to transmit. Coady (1992) treats testimony as a fundamental social epistemic practice that cannot be reduced to inference from non-testimonial premises. Goldman’s (1999) veritism connects credibility to truth-tracking and provides the philosophical scaffolding for asking what trust signals are actually evidence of. Lackey (2008) develops the conditions under which reliance on testimony is epistemically warranted. Fricker (2007) analyses testimonial injustice and surfaces the distortions introduced when credibility judgements are systematically biased by who the speaker is rather than what the speaker says. The literature is rigorous about evaluator positions, and it has not been translated into operational frameworks for entity trust evaluation in commercial or platform contexts.

2.3 Authority, reputation, and notability

The sociology of authority provides the conceptual scaffolding for treating Authority as a peer-conferred property distinct from Notability. Weber’s (1922/1978) classical typology distinguishes charismatic, traditional, and rational-legal authority. Bourdieu (1986) develops the concept of symbolic capital, which extends the analysis of authority into the dynamics through which standing accumulates and is converted across fields. Luhmann (1979) treats trust as a mechanism for the reduction of social complexity, a framing that recent scholarship continues to apply when distinguishing interpersonal trust from system trust (Kroeger, 2019); the present model operationalises one contemporary mechanism of that reduction, triangulation across Self, Peers, and Audience. As a practical instantiation of independent-secondary-source recognition, the Wikipedia notability guidelines (Wikipedia, n.d.-a, n.d.-b) operationalise referential salience in editorial practice; they serve here as an operational example rather than a theoretical foundation. Gillespie (2014) analyses how algorithmic authority is produced and legitimated. None of these literatures has been integrated with the source credibility tradition into a single framework that practitioners can apply.

2.4 Computational trust and algorithmic authority

The computational trust literature models trust as a quantifiable signal that propagates through networks. Castelfranchi and Falcone (2010) develop a socio-cognitive and computational model of trust that integrates cognitive states and quantitative reputation signals. Taddeo (2010) models trust between artificial agents specifically, separating direct trust from reputation-based trust in multi-agent systems. Pasquale (2015) analyses opaque algorithmic power in commercial and governmental information systems. Lewandowski (2012) bridges the search-engine literature with credibility theory, examining how ranking systems and credibility signals interact in web search. Gillespie’s (2014) treatment of algorithmic relevance applies directly to the question of how machine evaluators construct authority over the sources they retrieve and rank. Early web-ranking models such as PageRank (Brin & Page, 1998) and HITS (Kleinberg, 1999) are not models of trust in the human sense, but they show how machine systems can operationalise authority-like signals through link structure, and how easily authority, popularity, and centrality can be confused when the underlying behaviour is not decomposed. That conflation is a recurring failure mode in web-scale evaluation: popularity read as authority. The peer-axis decomposition developed in section 6 offers a conceptual response: deference (Authority) and reference (Notability) are different peer behaviours, and any measure that aggregates them inherits the confusion. This literature has informed the design of trust signals inside recommender systems and search ranking. The operational frameworks it has produced are platform-internal rather than entity-facing.

2.5 AI-mediated evaluation

The most recent body of work concerns how large language models, retrieval-augmented generation systems, and AI assistants evaluate sources. Bender, Gebru, McMillan-Major, and Shmitchell (2021) characterise large language models as systems that learn statistical patterns from web-scale corpora without grounded understanding, which provides the corpus-mediated basis for the Mirror Principle developed in Section 7. Mitchell et al. (2019) introduce model cards as a documentation practice for machine learning models, supporting Transparency as a property that applies to machine evaluators themselves and not only to the entities they evaluate. Ouyang et al. (2022) document the InstructGPT training procedure, including reinforcement learning from human feedback (RLHF) as a tuning mechanism that introduces a second-order human-evaluator signal into model behaviour. Lewis et al. (2020) introduce retrieval-augmented generation (RAG), in which the model retrieves from an external corpus at inference time, extending the corpus-mediation observation beyond pretraining into deployment. Gao et al. (2023) survey the RAG literature, including the credibility and traceability problems that arise when retrieved content is integrated into generated outputs. Ji et al. (2022) and Maynez et al. (2020) survey the hallucination and factuality literatures, documenting the conditions under which AI systems fabricate or misattribute, with direct implications for the Corroboration-versus-Repetition distinction developed in Section 8. Liu et al. (2023) document the “lost in the middle” phenomenon, in which information positioned in the middle of long contexts is systematically under-weighted, providing a useful caveat against assuming AI systems read all evidence equally.

2.6 Search quality and E-E-A-T

Google’s Search Quality Rater Guidelines (Google, 2023) formalise E-E-A-T (Expertise, Experience, Authoritativeness, Trustworthiness) as the principal trust-evaluation construct for human raters assessing search quality. The framework has been enormously influential in SEO practice and platform governance. It was never designed as a formal academic theory, and its conflation of evaluator positions has been an acceptable simplification for its original purpose. The provenance matters for practice. E-E-A-T is a rater-facing construct: it was written to give Google’s quality raters a vocabulary for evaluating content, not to give entities an operating procedure for improving their position. A vocabulary of evaluation describes what good looks like; it does not specify whose behaviour must change or through which lever. Practitioner effort spent optimising against the vocabulary, without a diagnostic structure underneath it, is effort without a sufficiently specified target. The NEEATT framework, an applied extension developed by Jason Barnard with Transparency contributed by Jarno van Driel and documented across the programme’s subsequent papers (Barnard, TKF-7), added Notability and Transparency to the E-E-A-T property set. The operational problem this paper addresses is what happens when practitioners and platforms now use E-E-A-T, or NEEATT, for diagnostic tasks the frameworks were not built to support.

Across these six literatures, the gap is consistent. Credibility theory, social epistemology, the sociology of authority, computational trust, AI-mediated evaluation research, and the search-quality literature each contribute pieces of the picture. None provides a single operational grid that separates evaluator positions while naming the properties each evaluator is competent to judge.

3. The conflation problem appears wherever experience and expertise are treated as the same property

The simplest illustration of the conflation problem is the rule everyone knows. Ericsson, Krampe, and Tesch-Römer (1993) showed that expert performance is produced by deliberate practice under feedback in a domain with measurable performance, accumulated over long periods. Popular discourse compressed this into the “ten thousand hours” rule, and the compression dropped the conditions that made the original finding hold. Most of what gets called ten thousand hours in popular usage is hours, sometimes under deliberate practice conditions, often not, and the result is a population of practitioners with plenty of experience and very different amounts of expertise distributed across the same headcount.

Experience and Expertise separate cleanly the moment you look at them. Experience is the accumulated time on task: years in the work, projects completed, problems encountered. Expertise is the quality of judgement that experience could produce and does not always produce: depth of understanding, the ability to see what less experienced practitioners miss, the capacity to make the call when the textbook does not apply. Expertise depends on what happens during the hours, not just on how many hours there are (Ericsson, Krampe, & Tesch-Römer, 1993).

Treat them as one property and the diagnostic question collapses. The practitioner wants to know which property to develop next, and the answer depends on whether the practitioner is short on time or short on judgement. A common pattern in career stagnation is being long on experience and short on expertise. Fast progress often comes from short, intense periods of deliberate practice that build expertise faster than time alone could. The rule that conflates the two cannot diagnose either condition.

This is the cleanest possible illustration of the broader problem. Operational trust discourse has often conflated distinct properties, and E-E-A-T retains this diagnostic limitation at the framework level. E-E-A-T names Experience and Expertise as separate properties, which is the right move, and it does the same conflation at a higher level by mixing them with Authoritativeness (a peer-conferred property) and Trustworthiness (an audience-conferred property) in a single flat list. E-E-A-T is not wrong about which properties matter. It is under-specified about whose judgement is doing the work of evaluating each one.

A more general statement of the problem follows. Any framework that presents trust as a list of attributes, without naming the actor whose judgement is constitutive of each attribute, loses the ability to predict what behaviour needs to change to move which property. The remaining sections develop the framework proposed to address that gap.

4. Trust is triangulated across three constitutive evaluator positions

Trust evaluation is distributed across three evaluator positions, each primarily responsible for judgements the others cannot fully constitute.

4.1 Formal definitions of the three actors

Self. The entity whose trust profile is being assessed. Capable of producing first-party evidence of its own properties (publications, disclosures, credentials, dated artefacts). Capable of intentional behaviour aimed at shaping how it is evaluated. The Self axis contains properties primarily demonstrated by the entity, even when their strongest validation comes from Peers or the Audience.

Peers. Domain-relevant evaluators with the competence or standing required to compare the entity against field norms (Weber, 1922/1978; Bourdieu, 1986; Rieh, 2002). Peers may be other practitioners, accredited bodies, journals, industry institutions, or any actor whose judgement carries weight inside the relevant field. Constitutive judge of properties that depend on intra-field evaluation: which entities the field defers to, which the field names, which the field agrees with.

Audience. Receivers, users, customers, readers, or systems that consume, rely on, or act upon the entity’s claims (Hovland, Janis, & Kelley, 1953; Hilligoss & Rieh, 2008). Constitutive judge of properties that depend on reception: confidence in the entity, recognition of the entity outside its niche, and the channel-side conditions through which the entity is or is not addressable.

4.2 Actor role is context-dependent and the constitutive role is stable

An institution can occupy more than one role across different contexts. A regulator may act as Peer when accrediting a practitioner, and as Audience when consuming a research finding. A publication may act as Peer when reviewing a paper, and as Audience when reporting on an industry. A search engine may act as Audience when consuming a source, and as Peer when ranking it against alternatives (Gillespie, 2014; Lewandowski, 2012).

The flexibility of role assignment across contexts is consistent with the framework’s structural claim, because what matters is which evaluator position is constitutive of which property in the context of the trust judgement being made. Context determines which actor is wearing which hat. The framework can still be applied, provided the actor’s role in the specific trust judgement is made explicit.

4.3 The three positions are independent and irreducible

The independence claim is what makes the framework operationally useful. The Self cannot constitute its own Notability, because Notability depends on reference by relevant others (Wikipedia, n.d.-a). The Self can create the conditions that make such reference more likely, and the property is not complete until others perform the referencing behaviour. Peers cannot grant Reach, because Reach depends on channels and infrastructures on the audience side. The Audience cannot give the Self Authority among practitioners, because Authority lives inside the Peer group and depends on what Peers say to each other (Rieh, 2002).

Mix the actor levels in a single list of properties and the diagnostic question collapses again, this time at the framework level. That is the operational hole the nine-cell grid closes.

4.4 Evidence can be observed across roles even when judgement is owned by one

A short methodological note is necessary to prevent a common objection. The actor assigned to each property is the actor whose judgement is constitutive of the property. It is not the only actor who can observe evidence of the property. Self may generate evidence of Experience (a dated publication record), and Peers may validate or contest that evidence (a peer review that confirms or disputes the dates or the substance). Peers may evaluate Authority through deference behaviour, and the Audience and machine evaluators may observe evidence of that peer deference through citation patterns, reference lists, and trade-press coverage (Athar & Teufel, 2012; Ding et al., 2014). The constitutive judgement and the observable evidence are different concepts. The framework assigns the constitutive judgement. Evidence travels widely.

5. Nine properties, three per actor

The grid is a decomposition of credibility: three actors across the top, three properties down each column. The nine cells are:

Self	Peers	Audience
Transparency	Authority	Trustworthiness
Experience	Notability	Celebrity
Expertise	Corroboration	Reach

What follows is the formal definition of each cell, the behaviour it measures, the kind of evidence that records it, and the lever through which it can be moved.

5.1 The Self axis

Transparency. Definition: the degree to which the entity discloses its identity, methodology, affiliations, and basis for its claims. Behaviour measured: publication under verifiable name, disclosure of affiliations and methods, citability of sources. Evidence type: author bylines, about-pages, disclosure statements, methodology sections, dated records of authorship. Improvement lever: deliberate disclosure practice, structured identity publication, transparent methodology documentation. Note: Transparency was contributed to NEEATT by Jarno van Driel. It is foundational because the other Self properties have nothing to attach to without a claimable identity behind them. Mitchell et al. (2019) extend the same principle to machine evaluators themselves through the model card practice, which is the Transparency analogue for the systems that evaluate other entities.

Experience. Definition: the accumulated time and exposure the entity has had in the relevant domain. Behaviour measured: years in the work, projects completed, problems encountered, sustained engagement over time. Evidence type: dated publications, projects, talks, work history, longitudinal records of participation in the field. Improvement lever: sustained activity over time. Experience accumulates at one hour per hour. It is the property least responsive to short-term effort.

Expertise. Definition: the quality of judgement the entity brings to work in the domain, beyond what time alone would produce (Ericsson, Krampe, & Tesch-Römer, 1993). Behaviour measured: depth of understanding, originality of contribution, capacity to handle non-routine cases. Evidence type: substantive publications, novel frameworks, demonstrated outcomes on difficult cases. Improvement lever: deliberate practice with structured feedback, originality of contribution to the field, depth of substantive work. Note on placement: Expertise sits on the Self axis because it is primarily demonstrated by the entity through substantive work. The strongest validation of Expertise typically comes from Peers (observable through the Authority they confer) and from Audience (observable through outcomes the entity produces for them). This is the general pattern: a property on the Self axis can be primarily demonstrated by the entity even when its strongest validation runs through the other two actors.

5.2 The Peers axis

Authority. Definition: the degree to which Peers defer to the entity’s judgement on questions in the field (Weber, 1922/1978; Rieh, 2002). Behaviour measured: deference behaviour by other practitioners (private consultation, citation as the source on contested questions, invitation to weigh in on difficult cases). Evidence type: citations in peer work, invited talks at peer events, expert testimony, advisory positions conferred by peers (Athar & Teufel, 2012). Improvement lever: repeated demonstration of competent judgement over time. Authority is built slowly and rarely transfers across fields.

Notability. Definition: the degree to which the entity is referred to by Peers when the topic is discussed. Behaviour measured: salience of the entity’s name in field discourse. Evidence type: mentions in trade press, references by peer practitioners, named appearances in field-level discussion, presence in industry award lists or canonical reading lists. The Wikipedia notability guidelines (Wikipedia, n.d.-a, n.d.-b) operationalise this property in editorial practice through the independent-secondary-source criterion. Improvement lever: visible contribution to field discourse, presence at field-level events, publication of work that becomes the reference point. Note: Notability was contributed to NEEATT by Jason Barnard. Notability differs from Authority: an entity can be the name everyone refers to (Notability) without being the judgement everyone defers to (Authority), and vice versa.

Corroboration. Definition: the degree to which independent voices align with the entity’s claims, framings, or positions. Behaviour measured: independent agreement by other competent voices, accumulation of confirming evidence from sources with material independence from the entity. Evidence type: independent publications citing and supporting the entity’s framing, peer commentary aligning with the entity’s claims, replication or confirmation of the entity’s findings (Shwed & Bearman, 2010). Improvement lever: being right early, being clear enough to cite, producing work other practitioners can verify and extend. Note: Corroboration is distinct from citation count, from mention count, and from consensus. Section 8 develops the four-way distinction in detail.

5.3 The Audience axis

Trustworthiness. Definition: the audience’s overall confidence that the entity can be relied on (Hovland, Janis, & Kelley, 1953; Hilligoss & Rieh, 2008). Behaviour measured: willingness of the audience to act on the entity’s claims, willingness to recommend the entity to others. Evidence type: reviews, testimonials, sustained client relationships, audience-side endorsements, repeat engagement. Improvement lever: delivered outcomes for the audience, sustained reliability, transparent handling of failures. Note: Trustworthiness, in this grid, means audience-perceived trustworthiness: the audience-side confidence that the entity can be relied upon. The property keeps its E-E-A-T name (Google, 2023) because continuity with the predecessor framework matters operationally, and it sits at the audience level, sharply distinguished from Peer-level properties such as Authority that E-E-A-T places in the same flat list. The definition does not deny that trustee-side qualities matter; the organisational trust literature models trustworthiness through ability, benevolence, and integrity (Mayer, Davis, & Schoorman, 1995; Schoorman, Mayer, & Davis, 2007). The grid assigns the constitutive judgement of reliance to the Audience position while allowing the evidence for that judgement to originate with Self and Peer signals.

Celebrity (used here as a technical term). Definition: the degree of recognition the entity has outside its specialist niche among general audiences, with broad public fame as the extreme form. Behaviour measured: recognition by audiences who do not themselves practise in the entity’s field. Evidence type: general-press coverage, public recognition, name salience in non-specialist conversation. Improvement lever: cross-domain visibility, work that bridges to wider audiences, sustained presence in general-audience channels. Note: Celebrity is the audience-side property that captures cross-niche recognition. It is distinct from Notability, which is recognition within a Peer group, and from Trustworthiness, which is audience confidence rather than recognition. In the model, Celebrity covers a spectrum from modest cross-niche visibility to general-audience fame. The colloquial sense of “celebrity” picks out the upper end of that spectrum.

Reach. Definition: the channel-level addressability of the audience: whether the entity can in principle get its message to the intended audience at all. Behaviour measured: size and quality of distribution channels, whether the entity has an operational route to the audience. Evidence type: distribution lists, channel partnerships, media access, platform presence, audience engagement metrics (Gillespie, 2014; Pasquale, 2015). Improvement lever: channel investment, audience development, platform partnerships. Note: Reach is assigned to the Audience axis because it names the audience-side condition under which trust judgements become operational: addressability. It is not a conscious audience judgement; it is the distribution condition that determines whether the relevant audience can be reached at all. The Audience axis therefore contains reception properties rather than three judgements of the same kind: confidence (Trustworthiness), recognition (Celebrity), and addressability (Reach). An entity with high Trustworthiness and Celebrity and no Reach is whispering in a soundproof room: the trust is real and operationally useless because the audience is not addressable.

5.4 E-E-A-T maps to four of the nine cells and NEEATT to six

The grid maps both predecessor frameworks into a more evaluator-specific structure without discarding their useful intuitions. In the terms of this model, E-E-A-T (Google, 2023) maps to four cells (Experience and Expertise on the Self axis, Authoritativeness as Authority on the Peers axis, Trustworthiness on the Audience axis). NEEATT (Barnard, TKF-7), developed by Jason Barnard with Transparency contributed by Jarno van Driel, added Notability on the Peers axis and Transparency on the Self axis, bringing the named count to six. The nine-cell grid adds Corroboration on the Peers axis and Celebrity and Reach on the Audience axis. The proposed grid offers a bounded operational specification under the paper’s three-actor assumption: each constitutive evaluator position receives three operationally distinct properties.

Cell	E-E-A-T	NEEATT	Nine-cell grid
Self / Transparency		✓	✓
Self / Experience	✓	✓	✓
Self / Expertise	✓	✓	✓
Peers / Authority	✓ (as Authoritativeness)	✓	✓
Peers / Notability		✓	✓
Peers / Corroboration			✓
Audience / Trustworthiness	✓	✓	✓
Audience / Celebrity			✓
Audience / Reach			✓

The two limitations the table records are distinct. E-E-A-T, by the grid’s measure, is incomplete (four of nine cells) and structurally flat (no property carries an evaluator assignment). NEEATT, the author’s own extension, closed part of the property gap and kept the flat structure: it added cells without adding evaluator specificity. The grid’s contribution addresses the second limitation, which neither predecessor explicitly resolves. The cells the grid adds are often the cells that explain why entities with strong E-E-A-T profiles nonetheless fail to translate that strength into outcomes. Corroboration explains why two practitioners with equal Authority are treated unequally. Reach explains why entities with strong Trustworthiness in their served audience cannot extend to new audiences. Celebrity explains why some entities punch above their Peer-level standing and others punch below.

6. The peer axis is the central conceptual contribution

Of the three columns, the Peers axis carries the contribution that most directly addresses the E-E-A-T conflation problem, and it is where credibility is most commonly misread in practice. E-E-A-T names one cell on this axis (Authoritativeness). NEEATT adds a second (Notability). The proposed peer-axis decomposition (Authority, Notability, Corroboration) names three properties that are operationally distinct, frequently misaligned, and routinely confused in practice. This is the conceptual centre of the paper, and the rest of the framework supports it.

6.1 Deference, reference, agreement

The three properties correspond to three distinct peer behaviours.

Authority is deference. Peers defer to the entity’s judgement when the question is contested or difficult (Weber, 1922/1978; Rieh, 2002). Authority shows up in private consultation, in citation as the source on a question, in invitation to settle disagreements between other practitioners. Authority accumulates through repeated demonstration of competent judgement over time, and it rarely transfers across fields.

Notability is reference. Peers refer to or recognise the entity as salient when the topic is discussed. Notability shows up in mentions, citations, named appearances in field discourse. Notability does not require that Peers agree with the entity. The named dissenter, whom everyone discusses but nobody agrees with, has high Notability and low Corroboration.

Corroboration is agreement. Independent voices align with the entity’s claims, framings, or findings (Athar & Teufel, 2012; Shwed & Bearman, 2010). Corroboration shows up in repeated independent confirmation, in replication, in commentary that extends or applies the entity’s framing. Corroboration is one mechanism through which contested claims may move toward consensus over time, provided the confirming voices are independent, competent, and substantively aligned.

6.2 The strategic positions on the peer axis

The three properties produce eight distinct strategic positions (each can score high or low, giving 2 to the power of 3 combinations). Five of those positions are common enough to deserve naming.

High Authority, low Notability. The quiet eminence. Peers defer privately, and the name does not surface in public field discourse. Common among technical experts in industries where Peer recognition runs through private channels rather than press visibility.

High Notability, low Authority. The famous commentator. Widely recognised by Peers as a name in the field, and not deferred to on difficult questions. Common among public-facing figures whose role is amplification rather than judgement.

High Corroboration, low Authority and low Notability. The diffuse consensus. The position is widely supported, the framing is widely repeated, and no single entity owns it. Common with claims that have become so generally accepted that the original source has faded from attribution.

High Authority, low Corroboration. The dissenting expert. Peers defer to the entity’s judgement in general, and the specific claim the entity is currently advancing is not yet (or no longer) widely agreed with. Common for entities advancing novel positions or contrarian readings.

High Notability, low Corroboration. The controversial figure. Widely referred to, widely discussed, not widely agreed with. Common for entities whose public profile depends on contested positioning.

The diagnostic value of these distinctions is that they prescribe different interventions. The quiet eminence needs Notability work (publication, field-discourse presence). The famous commentator needs Authority work (substantive contribution, demonstrated judgement). The dissenting expert needs Corroboration work (sustained argument until other voices align, or a tactical retreat if the claim cannot be defended). The single label “Authoritativeness” cannot make any of these prescriptions, because it collapses all three properties into one cell.

6.3 Notability without Corroboration is the trap

The peer-axis decomposition surfaces a common failure mode that E-E-A-T and NEEATT cannot diagnose. An entity invests heavily in Notability work (publication, field-discourse presence, naming new concepts, attending field events) and accumulates the visible markers of presence. The entity’s Corroboration, however, lags: independent voices are not aligning with the entity’s framings, even though the entity’s name is widely referred to.

The trap is that the visible markers of Notability look like progress. The trade press mentions the entity. Peers invite the entity to speak. The entity’s name appears in industry conversations. Inside the model, this is a single-axis success: one of three Peer cells is moving. The other two are not. The entity is becoming famous without becoming agreed with, and the AI-mediated information environment treats famous-without-agreed-with very differently from famous-with-agreed-with (Bender et al., 2021; Bikhchandani, Hirshleifer, Tamuz, & Welch, 2021).

The trap is the kind the framework is designed to make visible.

7. The Mirror Principle, in its defensible form

The framework is built as a model of human trust evaluation. Its extension to machine evaluators therefore requires a separate mechanism, formulated carefully so that it does not overclaim about machine cognition.

7.1 The principle as a mechanism statement

The Mirror Principle. Machine evaluators are constrained by the trust signals represented in training corpora, retrieval corpora, and human-feedback procedures. AI systems trained on, tuned by, or connected to human-generated information environments (Bender et al., 2021) are exposed to the trust-signal structures of those environments, and their outputs may reproduce, compress, amplify, or distort those structures depending on architecture, retrieval access, ranking logic, and feedback regime. Trust assessments produced by such systems tend to reproduce the broad three-actor structure (self-presentation, peer validation, audience reception) that organises human trust evaluation in the corpora the systems were trained on. The Mirror Principle is a falsifiable expectation about machine-evaluator behaviour, not an identity claim about machine cognition. The principle predicts outputs only where the relevant evidence is available to the system and where the system’s architecture, retrieval configuration, ranking logic, or training regime gives that evidence a path to influence the output.

The principle does not claim that machines reason about trust the way humans do. The claim is structural and mechanical. The training corpus is human-authored (Bender et al., 2021). The reward signals in reinforcement learning from human feedback are human-produced preferences (Ouyang et al., 2022). The retrieval corpora that large language models draw on at inference time through retrieval-augmented generation are overwhelmingly human-produced (Lewis et al., 2020; Gao et al., 2023). The trust signals the model can read, weight, and reproduce are therefore the trust signals embedded in human-produced material, which were themselves produced by humans for human-evaluator audiences operating in roughly the structure the nine-cell grid describes.

7.2 The principle as a research expectation

Stated as a falsifiable expectation, the Mirror Principle predicts the following pattern. When machine-mediated systems produce trust-relevant outputs (citation, recommendation, ranking, surfacing, summarisation, or exclusion), those outputs should correlate with the entity’s profile across the nine-cell grid where the system has access to the relevant evidence and is configured to use it. The correlation should be especially strong on the cells that depend on independent corroboration, third-party editorial coverage, and specific resolved outcomes, because those are the cells where human trust signals are most clearly encoded in publishable form.

The principle predicts the inverse correlation as well. Entities that score high on Self-axis disclosure and low on Peer-axis corroboration and Audience-axis recognition should be discounted by machine evaluators in the same direction human evaluators discount self-declaration unbacked by external signal. Entities that score high on Peer-axis Corroboration and low on Audience-axis Reach should appear in machine-mediated responses when the relevant queries are asked, and they will not surface broadly to audiences that are not asking the relevant queries.

7.3 The principle’s limits

The principle is a mechanism statement, not an identity claim. Machine evaluators differ from human evaluators in several respects that bound the principle’s reach.

Machine evaluators access more evidence than any individual human evaluator. They can integrate signals across a larger evidence base, which compresses some properties (Notability becomes easier to detect at scale) and amplifies others (Corroboration becomes detectable across a wider corpus than any single human reader could survey).

Machine evaluators are subject to training-data biases that human evaluators may not share (Bender et al., 2021). The corpus the model was trained on is not a uniform sample of the human-produced information environment. Entities under-represented in the training corpus will be evaluated with less precision, and the under-representation may itself be a function of the entities’ nine-cell profile in the period the corpus was assembled.

Machine evaluators do not weight all retrieved evidence equally. Liu et al. (2023) document the “lost in the middle” phenomenon, in which information positioned in the middle of long contexts is systematically under-weighted relative to information at the beginning or end. Position effects of this kind mean that the Mirror Principle’s correlations operate within engineering constraints that vary by architecture and inference-time configuration.

Machine evaluators are tuned by feedback loops (RLHF, evaluation suites, deployment metrics) that introduce second-order biases (Ouyang et al., 2022). The Mirror Principle predicts that the second-order biases also reflect human trust-evaluation patterns, because the human preferences that produced them were generated by humans operating within the same three-actor structure. The prediction is empirical, and it is testable.

Machine evaluators are also subject to hallucination and factuality failures (Ji et al., 2022; Maynez et al., 2020), which can produce trust assessments that do not correspond to any reading of the underlying corpus. The Mirror Principle predicts the correlations that hold in the absence of hallucination. The hallucination literature documents the conditions under which the correlations break.

The principle, in this defensible form, is a useful organising claim for the rest of the framework. The grid that describes human trust evaluation also describes the machine evaluator’s reading of the human-produced corpus, because the machine is structurally constrained to reproduce the trust-signal structures the corpus encodes.

8. Corroboration is the strategically compounding lever once thresholds are met

Of the nine cells, Corroboration is the cell whose absence most commonly explains why entities with strong Self-axis profiles fail to translate that strength into Peer-axis standing, and it is the cell that compounds most directly under deliberate operational effort once minimum thresholds of Transparency, Experience, Expertise, and Reach are in place.

8.1 Why Corroboration moves under deliberate action once prerequisites are met

The prerequisite condition matters. An entity with low Transparency cannot build Corroboration, because corroborating sources have nothing claimable to corroborate. An entity with low Expertise cannot build Corroboration, because the substantive material the corroborating sources would align with does not exist. An entity with low Reach cannot benefit from Corroboration in audience-side outcomes, because the corroborated position does not get to the audience. Corroboration is the strategically compounding lever in the configuration where the prerequisite cells are sufficient, and it is the lever least worth pulling when the prerequisites are absent.

Conditional on prerequisites, Corroboration moves under deliberate action in a way the other Peer-axis cells do not. Authority moves slowly through repeated demonstration of competent judgement, and the demonstrations have to be observed and accepted by Peers, which the entity cannot force (Rieh, 2002; Weber, 1922/1978). Notability moves with publication effort, and the move is bounded by what the field is willing to refer to. An entity can publish prolifically and not become Notable if the publications do not enter the discourse Peers are having.

Corroboration is different in this respect: it moves when independent voices choose to align with the entity’s framing, and the entity can influence that choice directly through a small set of actions. Being right early enough that the position has time to be confirmed. Being clear enough to be citable. Being the easiest source for the next person who wants to make a related argument. Producing artefacts that other practitioners can extend, replicate, or apply.

The argument is therefore qualified. Corroboration is not a universal first-priority lever. It is the cell that compounds most directly once the configuration permits it, which is the configuration most strong-profile entities arrive at well before they realise the Corroboration cell is the one limiting their Peer-axis growth.

8.2 Corroboration, consensus, citation, repetition

The conceptual care this section requires is that Corroboration is easily confused with three adjacent concepts. The confusion has practical consequences in AI-mediated environments, where machine evaluators that fail to distinguish between corroboration and repetition will systematically misweight evidence (Bender et al., 2021; Bikhchandani et al., 2021).

Corroboration is accumulating independent agreement. The agreement is real (the corroborating source endorses the claim, framing, or finding, on the basis of independent evaluation), the source is materially independent (the corroborating source has enough separate evidentiary basis, incentive structure, or evaluative agency that its alignment is not merely repetition, syndication, or coordinated amplification), and the corroboration accumulates over time as more sources arrive at the same conclusion (Shwed & Bearman, 2010).

Consensus is the mature state in which Corroboration has saturated and active disagreement has become marginal (Shwed & Bearman, 2010). Consensus is the downstream outcome of Corroboration. It is not the same property as Corroboration, and treating it as the same property loses the ability to distinguish positions that are in the process of being corroborated from positions that have already settled.

Citation is a visible trace of reference. Citation may or may not signal agreement (Athar & Teufel, 2012; Ding et al., 2014). A paper cited as the source of a claim is cited. A paper cited as an example of an error is also cited. Citation is a more general category than Corroboration, and the framework treats Citation as evidence that may indicate Corroboration, Notability, or Authority depending on the citation’s content. A retrieval system that treats citation count as a Corroboration proxy will misattribute, and the content-based citation analysis literature has shown how unreliable raw citation counts are as a proxy for agreement specifically (Athar & Teufel, 2012; Ding et al., 2014).

Repetition is mention without independent evaluation. Repetition occurs when sources cite each other without independently confirming the claim, when content gets quoted across many surfaces without any of the quoting sources having evaluated it, or when a claim becomes widely repeated through information cascades (Bikhchandani et al., 2021). Repetition is the failure mode that AI-mediated environments are particularly vulnerable to, because retrieval systems can surface repeated content at high volume without distinguishing repetition from independent confirmation. The hallucination and factuality literatures (Ji et al., 2022; Maynez et al., 2020) document related failure modes in which generated content propagates without grounding, which is the AI-era analogue of the same problem.

The four-way distinction is the analytical content of this section. The operational consequence is that practitioners building Corroboration must invest in independent voices arriving at the same conclusion through their own evaluation, rather than in volume of mention.

Material independence can be tested with four questions. Did the corroborating source evaluate the claim independently? Does the source have a reason to agree beyond its relationship with the entity? Does the source add evidence, application, critique, replication, or new context? Would the source plausibly maintain the claim if the entity disappeared? The last question is the sharpest separator: a claim that exists only because sources copy the entity is repetition, and a claim that survives without the entity because other sources can defend it on their own evidence is corroboration in formation.

9. The diagnostic application

The model’s operational value is the diagnostic application. Producing a nine-cell snapshot of an entity makes the bottleneck cell visible, and the bottleneck cell names the actor whose behaviour must change and the mechanism through which it can change. This section walks through five stylised cases.

9.1 The established academic with low public discoverability

Profile. Twenty-five years of substantive publication. Strong record of original contribution to the field. Cited in the field’s canonical reading lists. Privately consulted by senior practitioners on difficult cases. No general-press coverage. Modest audience-side reach outside the academic field.

Nine-cell snapshot. Self axis: Transparency high, Experience high, Expertise high. Peers axis: Authority high, Notability medium, Corroboration high. Audience axis: Trustworthiness high inside the served audience, Celebrity low, Reach low.

E-E-A-T diagnosis. The four E-E-A-T cells score high. E-E-A-T cannot identify a deficit, because the entity is strong on every cell E-E-A-T names. The E-E-A-T diagnosis is “high E-E-A-T, all is well.”

Nine-cell diagnosis. The bottleneck is on the Audience axis. The entity is invisible to audiences outside the served niche, despite being strong on every Self and Peer cell. The intervention is audience-side: Reach investment (channel partnerships, broader distribution), with secondary attention to Celebrity (cross-domain visibility) if the audience-side mission requires it.

9.2 The viral influencer with weak peer authority

Profile. Large general-audience following. Widely recognised by name outside any specialist field. Limited substantive publication. Limited Peer-side recognition.

Nine-cell snapshot. Self axis: Transparency medium, Experience low to medium (depending on tenure), Expertise low. Peers axis: Authority low, Notability medium (the entity is referred to, often skeptically), Corroboration low (peer voices do not align with the entity’s specialist claims). Audience axis: Trustworthiness medium (audiences trust the entity for entertainment value, not for substantive claims), Celebrity high, Reach high.

E-E-A-T diagnosis. The four E-E-A-T cells produce a mixed reading. Trustworthiness can be high in the served audience, Authoritativeness is low, Experience and Expertise are low. E-E-A-T correctly identifies a deficit, and it cannot prescribe which actor’s behaviour needs to change.

Nine-cell diagnosis. The bottleneck is on the Peers axis, specifically Authority and Corroboration. The intervention is Peer-facing: substantive contribution to the field that Peers can evaluate, sustained demonstration of competent judgement on field-relevant questions. Audience-side strength does not transfer to Peer-side standing without Peer-side work.

9.3 The niche practitioner with strong client trust and poor peer corroboration

Profile. Strong client outcomes. High audience-side trust. Limited industry visibility. Peer recognition is limited because the practitioner has not engaged sustainedly with field-level discourse.

Nine-cell snapshot. Self axis: Transparency medium, Experience high, Expertise high. Peers axis: Authority low, Notability low, Corroboration low. Audience axis: Trustworthiness high in served audience, Celebrity low, Reach medium.

E-E-A-T diagnosis. The four E-E-A-T cells produce a mixed reading. E-E-A-T does not specify whether the deficit lives on the Peer axis or on the Audience axis.

Nine-cell diagnosis. The bottleneck is the entire Peer axis. The intervention is Peer-facing work: publication of substantive material the field can evaluate, sustained presence in field-level discourse, accumulation of independent voices that align with the practitioner’s framings. The Self-axis strength is sufficient to support Peer-axis growth. The Peer-axis weakness is what limits the practitioner’s transfer of strong client-side outcomes into industry-level standing.

9.4 The dissenting expert with high authority and low corroboration

Profile. Deep substantive expertise. Recognised by Peers as a competent judge on contested questions. Currently advancing a position that other Peers are not yet (or no longer) aligning with.

Nine-cell snapshot. Self axis: Transparency high, Experience high, Expertise high. Peers axis: Authority high, Notability high, Corroboration low. Audience axis: variable.

E-E-A-T diagnosis. High E-E-A-T across the board. E-E-A-T cannot detect the Corroboration gap, because E-E-A-T does not separate Corroboration from Authoritativeness.

Nine-cell diagnosis. The bottleneck is Corroboration. The intervention has two strategic forks. Sustain the position long enough for Corroboration to accumulate, on the bet that the position is correct and other voices will arrive at the same conclusion (Shwed & Bearman, 2010). Or revise the position if Corroboration fails to accumulate after sustained effort, on the recognition that the deficit may be a signal that the position will not survive evaluation. The choice between forks is substantive, and the framework cannot make it. The framework can make the deficit visible, which is what E-E-A-T cannot do.

9.5 The corporate brand with high notability and low perceived trust

Profile. Widely recognised brand name. Visible in the trade press. Substantive products and services. Audience-side confidence has been damaged (regulatory issues, public scandals, sustained customer dissatisfaction).

Nine-cell snapshot. Self axis: Transparency variable (often deficient where the audience-side issues are unresolved), Experience high, Expertise high. Peers axis: Authority medium to high (the brand is taken seriously by industry), Notability high, Corroboration variable. Audience axis: Trustworthiness low, Celebrity high, Reach high.

E-E-A-T diagnosis. E-E-A-T identifies the Trustworthiness deficit. E-E-A-T does not prescribe the intervention because the diagnosis does not separate audience-side trust deficits from Self-axis Transparency deficits or from Peer-axis Corroboration deficits that may be contributing to the audience-side outcome.

Nine-cell diagnosis. The bottleneck is on the Audience axis (Trustworthiness), with likely contribution from the Self axis (Transparency). The intervention is the audience-side work that rebuilds confidence, which typically requires Self-axis Transparency work as a prerequisite (acknowledgement, disclosure, methodology change). Notability and Reach should not be invested in further until the Trustworthiness gap is closed, because broader Reach to a low-Trust position amplifies the audience-side damage.

9.6 What the comparison shows

The five cases share a structural feature. In each case, E-E-A-T either fails to identify the deficit or identifies it without naming the actor whose behaviour needs to change. The nine-cell grid identifies the deficit and names the actor, which converts the diagnosis into an intervention path. The operational improvement is the result of evaluator specificity rather than of new properties. The grid does not invent new trust concerns; it reorganises them by evaluator position, separating the cells the predecessor frameworks were already reaching toward.

9.7 The diagnostic operating sequence

The five cases follow a single operating sequence that converts the grid from description into intervention. Score each cell qualitatively (low, medium, high). Identify the lowest cell that blocks the intended outcome. Name the evaluator position that constitutes that cell. Select interventions aimed at changing that evaluator’s behaviour. Re-measure after a defined interval. Each step answers one diagnostic question: the weakest cell names the current trust bottleneck, the constitutive actor names whose behaviour must change, the cell’s evidence type names what to measure, and the cell’s improvement lever names what to do next.

10. Limitations and directions for empirical validation

The framework is bounded in six respects. These limits qualify the claim rather than nullify it.

Context dependency. Actor roles and property weights vary by domain. Medical trust, financial trust, academic trust, and entertainment trust weight the nine cells differently. The grid is designed to travel across domains: the three actors and the nine properties are proposed as a structurally stable decomposition, with domain-specific weighting. The weights with which the cells matter for any particular trust judgement are domain-specific, and the framework does not specify those weights. Empirical work that establishes domain-specific weightings is a natural next step.

Measurement risk. Proxies for the nine properties (citations as a proxy for Authority, mentions as a proxy for Notability, backlinks as a proxy for Corroboration, review counts as a proxy for Trustworthiness) can distort the underlying property if treated as the property itself (Athar & Teufel, 2012; Ding et al., 2014). The framework defines each property at the level of behaviour and judgement. The proxies are useful where they correlate with the underlying property, and they are misleading where the correlation breaks. Empirical applications of the framework must validate the proxies against the underlying properties, particularly in AI-mediated environments where proxy gaming is widespread.

Self-report and verifiability. The Self axis depends on what the entity discloses about itself. Transparency, Experience, and Expertise are partially verifiable through external evidence (dated publications, peer-validated credentials, demonstrated outcomes), and they are also subject to self-presentation bias. The grid’s evidence-type column names verifiable forms of evidence for each property. The strength of the verification depends on the rigor of the verification process.

Mirror Principle empirical status. The Mirror Principle is stated as a mechanism with empirical predictions. The predictions have not been systematically tested across multiple AI architectures, training corpora, and deployment contexts. The principle is consistent with current observations of AI evaluation behaviour and with the corpus-mediation, RLHF, RAG, and hallucination literatures (Bender et al., 2021; Ouyang et al., 2022; Lewis et al., 2020; Gao et al., 2023; Ji et al., 2022; Maynez et al., 2020). Systematic empirical work would test the predicted correlations between nine-cell profiles and machine-evaluator outputs across model families. Initial cross-model citability measurement in industry settings, in which weighted citability was scored for the same entity across nine AI models, is consistent with the predicted direction (Authoritas, 2025); systematic validation across entities, architectures, and contexts remains open.

Operational use versus descriptive use. The framework can be used to describe trust evaluation as it occurs, or to prescribe operational interventions to improve trust outcomes. The descriptive use makes few assumptions. The prescriptive use assumes that the deliberate actions named in each cell’s improvement-lever entry actually move the underlying property in the expected direction. The assumption is plausible, and it is not proven for every cell in every domain.

Empirical work that would validate the model. Three programmes of work would convert the conceptual framework into an empirically grounded one. Longitudinal case studies that score entities across the nine cells over time and observe trust outcomes (citation patterns, recommendation by AI systems, audience-side conversion). Controlled comparisons that test diagnoses produced by E-E-A-T, NEEATT, and the nine-cell grid against independent ground truth on the same entity set. Machine-evaluator experiments that vary training corpora, retrieval configurations, and feedback regimes, and measure the predicted Mirror Principle correlations across them.

11. Conclusion

In the terms of this model, E-E-A-T maps to four of the nine cells, and NEEATT can be mapped to six. The proposed grid offers a bounded operational specification under the paper’s three-actor assumption, and the three cells the grid adds are the cells that most directly explain why entities with strong predecessor-framework profiles nonetheless fail to translate that strength into outcomes. Corroboration explains why two entities with equal Authority diverge in industry adoption. Reach explains why entities with strong Trustworthiness in their served audience cannot extend to new audiences. Celebrity explains why some entities punch above their Peer-level standing and others punch below.

The contribution of the paper is not the discovery of new properties. The contribution is evaluator specificity: the assignment of each property to the actor whose judgement is constitutive of it, and the consequent ability to diagnose a trust deficit by identifying the actor whose behaviour must change. E-E-A-T’s properties are mapped in, NEEATT’s contributions are mapped in, the three new cells are named, and the operational diagnostic that the predecessor frameworks could not deliver becomes available.

The peer-axis decomposition is the conceptual centre of the framework. Authority is deference, Notability is reference, Corroboration is agreement. The three are independent, frequently misaligned, and routinely confused. Distinguishing them surfaces strategic positions and operational interventions that single-label frameworks cannot reach.

The Mirror Principle, in the defensible form developed in Section 7, extends the framework to machine evaluators. Machine evaluators are constrained by the trust signals represented in training corpora, retrieval corpora, and human-feedback procedures (Bender et al., 2021; Ouyang et al., 2022; Lewis et al., 2020). The nine-cell grid that models human trust evaluation therefore predicts patterns in machine-evaluator behaviour at the structural level. Empirical work to test the predictions across architectures and contexts is required; initial cross-model citability measurement in industry settings is consistent with the predicted direction (Authoritas, 2025).

The proposed grid offers an operational specification of the trust-evaluation problem that E-E-A-T identifies and does not structurally decompose. The grid decomposes credibility into implementable units: each cell carries a named evaluator, a definition, an evidence type, and an improvement lever. Nine cells, three actors, one diagnostic. The trust discourse moves from descriptive to operational when each cell is named at the evaluator position constitutive of it. Whatever decomposition a practitioner prefers, the layer is the same: credibility, evaluated across Self, Peers, and Audience.

AI assistance disclosure

The author used large language model assistants (Anthropic Claude and Perplexity) during the preparation of this paper, for drafting support, editorial critique, structural review, and reference checking. The conceptual model, the three-actor structure, the peer-axis decomposition, the Mirror Principle, and all framework claims are the author’s. The author reviewed every section and takes full responsibility for the final text and the references. The disclosure follows the Transparency property the paper itself defines: the basis of the work is disclosed so that evaluators, human and machine, can weigh it.

References

Athar, A., & Teufel, S. (2012). Context-enhanced citation sentiment detection. Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 597-601. https://aclanthology.org/N12-1073/

Authoritas. (2025). Can you fake it ’til you make it in the age of AI search? Authoritas. https://www.authoritas.com/blog/can-you-fake-it-til-you-make-it-in-the-age-of-ai-search

Barnard, J. (TKF-7-20364742). AI-era commercial architecture: A survey and synthesis of business strategy, marketing transformation, and algorithmic intermediation, 2018-2026. Working paper, Kalicube. Zenodo. https://doi.org/10.5281/zenodo.20364742

Barnard, J. (TKF-10-20364725). AI-Era Business Engineering: The Integrating Frame for Commercial Architecture in the Age of Algorithmic Intermediation. Working paper, Kalicube. Zenodo. https://doi.org/10.5281/zenodo.20364725

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623. https://dl.acm.org/doi/10.1145/3442188.3445922

Bikhchandani, S., Hirshleifer, D., Tamuz, O., & Welch, I. (2021). Information cascades and social learning. Working paper. https://arxiv.org/abs/2105.11044

Bourdieu, P. (1986). The forms of capital. In J. G. Richardson (Ed.), Handbook of theory and research for the sociology of education (pp. 241-258). Greenwood. https://home.iitk.ac.in/~amman/soc748/bourdieu_forms_of_capital.pdf

Brin, S., & Page, L. (1998). The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems, 30(1-7), 107-117. https://doi.org/10.1016/S0169-7552(98)00110-X

Castelfranchi, C., & Falcone, R. (2010). Trust theory: A socio-cognitive and computational model. Wiley. https://onlinelibrary.wiley.com/doi/book/10.1002/9780470519851

Coady, C. A. J. (1992). Testimony: A philosophical study. Clarendon Press. https://academic.oup.com/book/25355/chapter/192425027

Ding, Y., Zhang, G., Chambers, T., Song, M., Wang, X., & Zhai, C. (2014). Content-based citation analysis: The next generation of citation analysis. Journal of the Association for Information Science and Technology, 65(9), 1820-1833. https://yingding.ischool.utexas.edu/Publication/CCA-final.pdf

Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363-406. https://graphics8.nytimes.com/images/blogs/freakonomics/pdf/DeliberatePractice(PsychologicalReview).pdf

Fogg, B. J. (2003). Persuasive technology: Using computers to change what we think and do. Morgan Kaufmann. https://dl.acm.org/doi/abs/10.5555/2821581; https://shop.elsevier.com/books/persuasive-technology/fogg/978-1-55860-643-2

Fricker, M. (2007). Epistemic injustice: Power and the ethics of knowing. Oxford University Press. https://academic.oup.com/book/32817

Gao, Y., Xiong, Y., Gao, X., Jia, K., Pan, J., Bi, Y., Dai, Y., Sun, J., Wang, M., & Wang, H. (2023). Retrieval-augmented generation for large language models: A survey. https://arxiv.org/abs/2312.10997

Gillespie, T. (2014). The relevance of algorithms. In T. Gillespie, P. J. Boczkowski, & K. A. Foot (Eds.), Media technologies: Essays on communication, materiality, and society (pp. 167-193). MIT Press. https://direct.mit.edu/books/edited-volume/3021/chapter/82555/The-Relevance-of-Algorithms; https://www.microsoft.com/en-us/research/wp-content/uploads/2014/01/Gillespie_2014_The-Relevance-of-Algorithms.pdf

Gillespie, T. (2018). Custodians of the internet: Platforms, content moderation, and the hidden decisions that shape social media. Yale University Press. https://archive.org/details/Custodians_of_the_Internet_by_Terleton_Gillespie

Goldman, A. (1999). Knowledge in a social world. Oxford University Press. https://academic.oup.com/book/32822

Google. (2023). Search quality evaluator guidelines. Retrieved from https://services.google.com/fh/files/misc/hsw-sqrg.pdf. See also https://developers.google.com/search/blog/2023/11/search-quality-rater-guidelines-update

Hilligoss, B., & Rieh, S. Y. (2008). Developing a unified framework of credibility assessment: Construct, heuristics, and interaction in context. Information Processing & Management, 44(4), 1467-1484. https://www.sciencedirect.com/science/article/abs/pii/S0306457307002038

Hovland, C. I., Janis, I. L., & Kelley, H. H. (1953). Communication and persuasion: Psychological studies of opinion change. Yale University Press. https://archive.org/details/communicationper0000unse; https://books.google.com/books/about/Communication_and_Persuasion.html?id=vil5tgAACAAJ

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2022). Survey of hallucination in natural language generation. ACM Computing Surveys. https://arxiv.org/abs/2202.03629

Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environment. Journal of the ACM, 46(5), 604-632. https://doi.org/10.1145/324133.324140

Kroeger, F. (2019). Unlocking the treasure trove: How can Luhmann’s theory of trust enrich trust research? Journal of Trust Research, 9(1), 110-124. https://doi.org/10.1080/21515581.2018.1552592

Lackey, J. (2008). Learning from words: Testimony as a source of knowledge. Oxford University Press. https://philpapers.org/rec/LACLFW-2

Lewandowski, D. (2012). Credibility in web search engines. https://arxiv.org/abs/1208.1011

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33. https://arxiv.org/abs/2005.11401

Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023). Lost in the middle: How language models use long contexts. https://arxiv.org/abs/2307.03172

Luhmann, N. (1979). Trust and power. Wiley.

Mayer, R. C., Davis, J. H., & Schoorman, F. D. (1995). An integrative model of organizational trust. Academy of Management Review, 20(3), 709-734. https://doi.org/10.5465/amr.1995.9508080335

Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 1906-1919. https://aclanthology.org/2020.acl-main.173/

Metzger, M. J. (2007). Making sense of credibility on the Web: Models for evaluating online information and recommendations for future research. Journal of the American Society for Information Science and Technology, 58(13), 2078-2091. https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.20672

Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., Spitzer, E., Raji, I. D., & Gebru, T. (2019). Model cards for model reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency, 220-229. https://arxiv.org/abs/1810.03993

Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35. https://arxiv.org/abs/2203.02155

Pasquale, F. (2015). The black box society: The secret algorithms that control money and information. Harvard University Press. https://digitalcommons.law.umaryland.edu/books/96/

Rieh, S. Y. (2002). Judgment of information quality and cognitive authority in the Web. Journal of the American Society for Information Science and Technology, 53(2), 145-161. https://ideas.repec.org/a/bla/jamist/v53y2002i2p145-161.html

Schoorman, F. D., Mayer, R. C., & Davis, J. H. (2007). An integrative model of organizational trust: Past, present, and future. Academy of Management Review, 32(2), 344-354. https://doi.org/10.5465/amr.2007.24348410

Shirky, C. (2008). Here comes everybody: The power of organizing without organizations. Penguin. https://books.google.com/books/about/Here_Comes_Everybody.html?id=9eCopwAACAAJ

Shwed, U., & Bearman, P. S. (2010). The temporal structure of scientific consensus formation. American Sociological Review, 75(6), 817-840. https://journals.sagepub.com/doi/abs/10.1177/0003122410388488

Sundar, S. S. (2008). The MAIN model: A heuristic approach to understanding technology effects on credibility. In M. J. Metzger & A. J. Flanagin (Eds.), Digital media, youth, and credibility (pp. 73-100). MIT Press. https://www.issuelab.org/resource/the-main-model-a-heuristic-approach-to-understanding-technology-effects-on-credibility.html; https://www.issuelab.org/resources/835/835.pdf

Taddeo, M. (2010). Modelling trust in artificial agents, a first step toward the analysis of e-trust. Minds and Machines, 20(2), 243-257. https://dl.acm.org/doi/10.1007/s11023-010-9201-3

Weber, M. (1978). Economy and society: An outline of interpretive sociology (G. Roth & C. Wittich, Eds.). University of California Press. (Original work published 1922.) https://books.google.com/books/about/Economy_and_Society.html?id=MILOksrhgrYC

Wikipedia. (n.d.-a). Wikipedia: Notability. https://en.wikipedia.org/wiki/Wikipedia:Notability

Wikipedia. (n.d.-b). Wikipedia: Notability (people). https://en.wikipedia.org/wiki/Wikipedia:Notability_(people)

Author

Jason Barnard is CEO and founder of Kalicube, a Digital Brand Intelligence™ consultancy based in France. His work focuses on brand intelligence, entity understanding, and AI-era business engineering. He coined Brand SERP (2012) and Entity Home (2015), and named each successive stage of the optimisation discipline’s evolution: Answer Engine Optimisation (AEO, 2017), AI Assistive Engine Optimisation (AIEO, 2024), and Assistive Agent Optimisation (AAO, 2025). He is the inventor on 17 pending patent applications (INPI) related to the diagnostic methodologies used in Kalicube’s platform. In the independent Authoritas Weighted Citability Score study (December 2025) he ranked first of more than 500 search professionals for citability across nine AI systems (https://www.authoritas.com/case-study-kalicube/). He has delivered Assistive Agent Optimisation training to enterprise teams in Google’s SEA Labs programme, and has keynoted on the discipline for international business audiences. His nineteen-article AI Authority Series runs on Search Engine Land from February to June 2026 (https://searchengineland.com/author/jason-barnard). ORCID: https://orcid.org/0009-0000-0353-6642

Working paper, v1.4, June 2026. Comments welcome to Jason Barnard at Kalicube.